At National Institute of Technology, I applied transfer learning with pre-trained CNNs to 3,300 ISIC dermoscopy images. I also built an autoencoder-based model with classical classifiers and compared the CNN backbones, identifying VGG16 as the strongest feature extractor.
On my medical Q&A fine-tuning project, I raised ROUGE-L from 0.16 to 0.30 by fine-tuning Qwen2.5-3B with 4-bit QLoRA on deduplicated MedQuAD pairs. I also improved the safety-test pass rate from 3/12 to 12/12 using intent routing, fixed emergency replies, and a dose-number filter.
I added retrieval-augmented generation over NIH Q&As with a grounding check, blocking ungrounded medical terms in 7 of 50 answers. For my RAG Study Assistance project, I built a chatbot over course notes and past exam papers with cited answers and filtered retrieval.
I'm studying Computer Science and Engineering at GIET University. My work spans machine learning and language-model applications using Python, PyTorch, and TensorFlow.

